LLM Basics 1 / 1 Connecting…

LLM Basics

Matteo Gala

“This technology will make us forget everything.”

Plato, Phaedrus · about 370 BC

Socrates. About writing.

“Not truth, but the semblance of truth.”

Where knowledge lives

“We’ll forget”Socrates, ~370 BC
Writing
Philosophy and lawPlato to Justinian
“Copy by hand”Trithemius, 1492
Printmovable type, ~1450
Shared scienceVesalius, 1543
“Too many books: confusing”Gessner, 1545
Booksmass editions, 1500s
DemocraciesUSA 1776, France 1789
“Won’t replace your newspaper”Stoll, 1995
Internet
Open knowledgeWikipedia, open courses
You are here
“It thinks for us”us, 2026
AI
?
?
Someone, one day, reading our op-eds

Fear sees the cost.
The gain comes later.

How learning changes

My mother’s school booklet

Ready-made sumsprimary school, 1950s
Calculatoruseless in a few years
A page of an 18th-century ready reckoner: columns of quantities and prices already computed
The answers, already computedA ready reckoner page, England, 18th century. Public domain, via Wikimedia Commons.

What did your grandparents learn at school that you never had to?

Finding 好 in a paper dictionary

1 · Radical index

2 strokes人 刀 力
3 strokes大 女 子 山
4 strokes心 手 日

女 → list on p. 52

2 · Characters under 女

+2 strokes奶 奴
+3 strokes如 她 好 妈
+4 strokes妙 妥 妨

好 → p. 183

3 · Page 183

好 hǎo

good; well; fine

Page numbers are illustrative
One of the two racks of IBM Deep Blue at the Computer History MuseumDeep Blue, Computer History Museum · Anton Chiang, CC BY 2.0

Chess, and then languages?

1997: a machine beats the world champion. We still play, for the mind.
Machines translate. Will we still learn languages, and why?
Watch: AlphaGo (2017), free on YouTube
The calculator retired one booklet. Which of ours will AI retire?
If an essay no longer proves you know, what will?
Which will become our chess, kept for the mind?
To answer, look under the hood.
Socrates: text can’t answer back
An LLM: the first text that answers

It reads tokens, not words

unbelievable
strawberry
Sustainability
ChatGPT
2026
Demo · platform.openai.com/tokenizerillustrative splits

It guesses the next piece

The capital of Italy is  

Rome0.92 a0.04 Milan0.02
illustrative numbers pick, append, repeat

Meanings become points

catdoghorseRomeParisMilanhappyglad animalscitiesfeelings

Close in meaning, close in space.

king − man + woman
≈ queen

word2vec, 2013

2-D illustration of thousands of dimensions

Plausible ≠ true

Stochastic parrotrecombines text
World modelbuilds a picture
Nobody knowsyet

So: verify

Hallucinationsconfidently wrong
Verifyanything you would sign

Context is its only memory

Total amnesiareads the desk every time
New chatempty desk
Too fullloses track
Whose data are you putting on the desk?

Same engine, different cars

The modelthe engine: predicts the next piece
A chat app+ hidden instructions, memory
A search assistant+ web search
An agent in your folder+ your files, programs, hours of work

The myth

Prompt engineering

Magic phrasessmall gains today
Contextmost of the result

Same request, two desks

PromptYou are an expert consultant. Write a great email about the delay.
ContextSmall hotel, formal owner, two years together. Launch slips 1 → 15 March: their photos came late. Stay warm, don’t apologise. 120 words, Italian.

Explain the job
as you would to a person.

Chat or workspace?

Chatknows what you typed
Workspacereads your files, runs programs
Live: “Make me a quiz on our Green transition module.”

Meanwhile, the agents…

Coordinator agent: “Matteo wants us to build a new module in the game”On its desk: the portal calendar and a log of what is done. It calls the agents below; it never reads a transcript.
1
Download agent

On its desk: the lesson title and date, the course portal

→ the lesson audio

2
Transcribe agent

On its desk: one audio file, nothing else

→ the transcript

3
Course notes agent

On its desk: the transcript and the notes written so far

→ notes, with what the professor took for granted filled in

4
Build game city agent

On its desk: the notes of every lesson in the module

→ a playable city of the study game

5
Translate agent

On its desk: the finished city

→ the city in Spanish, Russian and French too

Not a better prompt: an ecosystem of agents, each with its own desk.

Don’t polish the prompt.
Fill the desk.

Over to you

Handout below the slides, same link.

This page follows the slides live: nothing to click. On a phone, turn it sideways for bigger slides.